Analyzing the Impact of Data Augmentation on the Explainability of Deep Learning-Based Medical Image Classification

Deep learning models are widely used for medical image analysis and require large datasets, while sufficient high-quality medical data for training are scarce. Data augmentation has been used to improve the performance of these models. The lack of transparency of complex deep-learning models raises ethical and judicial concerns inducing a lack of trust by both medical experts and patients. In this paper, we focus on evaluating the impact of different data augmentation methods on the explainability of deep learning models used for medical image classification. We investigated the performance of different traditional, mixing-based, and search-based data augmentation techniques with DenseNet121 trained on chest X-ray datasets. We evaluated how the explainability of the model through correctness and coherence can be impacted by these data augmentation techniques. Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) methods were used. Sanity checks and overlapping scores were applied to confirm the correctness and coherence of explainability. The results indicate that both LIME and SHAP passed the sanity check regardless of the type of data augmentation method used. Overall, TrivialAugment performs the best on completeness and coherence. Flipping + cropping performs better on coherence using LIME. Generally, the overlapping scores for SHAP were lower than those for LIME, indicating that LIME has a better performance in terms of coherence.